A new algorithm for transforming pixel art into scalable vector images has been published in a joint paper from Seattle-based Microsoft Research and the Hebrew University of Jerusalem.
“Depixelizing Pixel Art” by Johannes Kopf and Dani Lischinski outlines an upscaling process that is suited to transforming pixel art of the like found in the low-resolution sprites of old console and PC games.
The algorithm interprets pixels in old assets as inserted by artists intentionally, and that lone pixels are not leftovers of a digital imaging process. It assesses pixel connections through the entirety of images to find out if individual pixels are part of greater shapes that need smoothing out or edges that need sharpening.
A series of heuristics are also included in the process to decipher complex pixel patterns and decide if they should be grouped or separated into distinct lines.
The team behind the paper has said that it hopes the algorithm will be usable in real-time emulators and as a basis for a process that generates additional frames to stop jumping in upscaled sprites in the future.
The paper presents example images of the process in action, with clear and distinct improvements on existing algorithms visible. Its does not, as accepted by the researchers, work for anti-aliased sprites however, and particularly angular assets tend to come out badly.
The full PDF can be found here.